> Le-Nguyen, Mihn-Huong. _Online ML-Based Predictive Maintenance for the Railway Industry_. # Online ML-based predictive maintenance for the railway industry - Maintenance is critical for the good function of the railway system - Corrective → too late, predetermined → too expensive - Solution: Condition-based predictive maintenance - Railway systems produce data streams - Offline (time-consuming, stable, cannot adapt to drifts) v Online (continuous, fast, unstable, adaptive) > RQ: Could we use online ML to achieve satisfacctory results for PdM of complex railway systems? - Cycle extraction → feature learning → health detection → prognostics ## 1. Cycle extraction - Change point detection is the only method than can handle multivariate data - A cycle must represent a function of a system → need for human knowledge → active learning using uncertainty sampling ## 2. Feature learning - LSTM auto-encoder ## 3. Health detection - Fault detection / identification / isolation - To prevent fault, detect _anomalies_ - Build a health score s.t. the evolution of any anomaly is explicitly monitored and its impact varies w.r.t. its severity - Static v (temporal) decaying versions